Research topics

Completed

Master’s

• An application for formal assessment and comparison of ontology visualizations
• A library for ontology learning based on unsupervised automated machine learning
• Automation of machine learning pipeline design based on the ontological approach
• A method for optimizing the number of clusters for the k-means algorithm based on Bayesian optimization
• A method for neural architect search for clustering problems
• An application for semantic-oriented visualization of ontologies based on the Logic Graphs language
• An application for extracting requirements bases from industry standards and classifiers
• Informed Object Detection for Computer Games
• Developing a mobile application for plant recognition using a knowledge base

PhD

• The method of diagrammatic verification of ontologies based on ALC description logic

In progress

Master’s

• Semantic Representation of Scientific Claims for Automated Evidence-based Verification
• Generation of NLU Benchmarks from Formal Representations for Contamination-Proof LLM Evaluation

Thesis suggestions:

Reinforcement Learning from Symbolic Reasoning Feedback: A Use Case from Building Compliance Checking

LLMs are trained using human feedback. However, human trainers are costly, and LLMs still struggle with reasoning-intensive tasks. If we describe such tasks formally, their solutions can be validated using symbolic reasoning engines, which can also provide explanations. These explanations, in turn, can serve as feedback to correct the outputs generated by LLMs. As a use case, we consider the task of building compliance checking and aim to train LLMs to translate natural-language requirements into formal representations.

Facilitating OWL DL Representation of Scientific Claims Using DeLTA for Consistency Checking

Consistency checking of scientific claims is essential for ensuring the coherence, reliability, and reproducibility of research. As the volume of scientific publications continues to grow, systematic methods for detecting and resolving inconsistencies can significantly enhance the quality and trustworthiness of scholarly communication. In this work, we apply Description Logic-based Text Annotation (DeLTA) to facilitate the representation of scientific claims in the OWL DL language. Once scientific claims are represented in OWL DL, their consistency can be automatically verified using logical reasoners.

Automating Translation of Natural Language Requirements into OWL DL Language Using DeLTA Annotation Schema: A Use Case from Building Compliance Checking

Enriching ontologies with complex axioms enables inference, querying, and consistency checking but is time-consuming and error-prone. Domain experts often lack ontology skills, while ontology engineers lack domain knowledge. DeLTA bridges this gap by mapping natural language directly to OWL DL, allowing domain experts to annotate text that can be automatically converted into formal axioms. In the construction domain, processing large volumes of building regulations and standards is essential for automated compliance and decision support. This work creates a benchmark corpus of DeLTA-annotated building requirements and trains NLP models to predict annotations and generate OWL DL representations automatically.

GraphReview: Structured Representation for Formal-Argumentation-Based Computational Support in Scientific Peer Review

The review process for scientific publications is becoming increasingly complex due to the rapid growth in submission volumes. We address these issues by conceptualizing peer review as an argumentative dispute between the authors of a manuscript and the reviewers evaluating it. We formalize this process using abstract argumentation frameworks and apply argumentation solvers to resolve such disputes. In this work, we will focus on developing user interfaces that facilitate the structured input of peer reviews.

Meta-learning for Data-Driven Retrieval of Relevant Scientific Literature

Due to the rapid growth in scientific publications, it has become increasingly challenging to identify research relevant to a given dataset. In this work, we apply methods from AutoML and meta-learning to facilitate the retrieval of relevant literature. Specifically, the tasks include developing methods and software for extracting meta-features of datasets, creating a knowledge base that links datasets with research papers, and developing methods for retrieving relevant papers using dataset meta-features.

Aligning Automatically Generated and Crowdsourced Scholarly Knowledge Graphs: A Case Study of CS-KG and ORKG

In this work, we focus on two compute-science-related knowledge graphs: CS-KG and ORKG. Because CS-KG is generated by an automated NLP pipeline and ORKG by crowdsourcing, the same entity may appear in both graphs with different URIs. Aligning these entity representations can be done using methods such as string or neighbor matching, or similarity based on text and graph embeddings, some of which are implemented in the OntoAligner Python library. The resulting mappings can be published as digital objects following FAIR principles and standards like SSSOM.

Other possible directions:

• Entropy adjustment as a metaheuristic for balanced clustering
• Entropy-based objective priors for Bayesian optimisation in centroid clustering
• Information-theoretic measures for computational knowledge extraction evaluation
• Diagrammatic reasoning over ontologies
• Visualization of semantic requirements for building model compliance